第103回統計地震学セミナー / The 103rd Statistical Seismology Seminar (Hybrid)
- 【Date＆Time】
- 23 October 2025 15:00-17:00
Admission Free, No Booking Necessary - 【Place】
- online (with zoom) and 4F Lounge (ISM)
- Zoom meeting:
https://us06web.zoom.us/j/8304503874?pwd=TkdxT1Z0YXFFWUt1c013eHkydTU2Zz09&omn=82778474047
Meeting ID: 830 450 3874
Passcode: 55338532 - 【Speaker1】15:00-15:40
- Linhai Wang
(Institute of Geophysics, China Earthquake Administration)
"Yearly-scale deep crustal fluid transfer implied by gravity changes from in-situ observations around the 2021 Yangbi Ms6.4 earthquake, China"
- Abstract: Gravity changes resulting from deep geo-fluid mass transfer can be detected through in-situ repeated observations. About three years prior to the 2021 Yangbi Ms6.4 earthquake, in China, two regions northeast and southeast of the epicenter exhibited significant gravity increases. For modeling this gravity change, we adopt an equivalent mass source model with multiple disks, together with a Monte Carlo-based inversion algorithm to estimate model parameters. Tests on synthetic data demonstrate that this approach simultaneously estimates geometric parameters and density change of the disk models, even in the presence of noise contamination. We applied this method to invert a double-disk model to approximate the mass "hypocentroids" based on the gravity increases in two regions. Our analysis identified fluid diffusion footprints reflected by background seismicity migration, with estimated rates of 5 m²/s and 3 m²/s, aligning with areas of significant gravity increase. This study suggests that deep geo-fluid migration began at least three years before the Yangbi earthquake. The findings reflect the existence of trapped fluids located to the northeast and southeast of the epicenter.
- 【Speaker2】15:40-16:20
- Jian Piao
(PhD student, Peking University)
"Modeling Long-term Seismic Hazard in North China Using ETAS-Simulated Catalogs and Ground-Motion Attenuation"
- Abstract: This study applies the Epidemic-Type Aftershock Sequence (ETAS) model to simulate long-term earthquake catalogs, which are integrated with region-specific ground-motion attenuation relationships to assess seismic hazard in North China over the next 50 years. By accounting for earthquake clustering, the ETAS framework provides time-dependent hazard estimates beyond the assumptions of the traditional Poisson-based SHA. The analysis shows that these temporal effects can substantially modify long-term hazard levels, particularly under conditions of high exceedance probabilities or multiple exceedances, underscoring the importance of incorporating time dependence into future SHA analyses.
- 【Speaker3】16:20-17:00
- Chengxiang Zhan
(PhD student, China University of Geosciences (Beijing))
"Negligible Magnitude Dependence in Real Seismicity"
- Abstract: Whether successive earthquake magnitudes are correlated remains debated. We develop a framework based on per-event information gain (PEIG) between a history-dependent modulated neural network (MNN) and a history-independent neural network (NN). This frameworks offer a new interpretation of a recently proposed information-gain measure, showing how the modulated architecture isolates the contribution of past magnitudes. That is, a significant positive LL gain suggests that past magnitudes enhance the predictability of the magnitude of the corresponding event, revealing magnitude dependence. Using synthetic catalogs, we show that the MNN not only detects genuine correlations when present, but also learns spurious 'dependence' induced by short-term aftershock incompleteness (STAI) and missing background events. For the real earthquake catalogs from California, the observed LL gains can be attributed to the STAI and the missing small events, rather than genuine physical correlations. Furthermore, when extending the analysis to positive magnitude differences of subsequent earthquake pairs, the dependence is either found. Overall, neither earthquake magnitudes nor their positive differences exhibit significant physical correlations even though they are examined with flexible deep neural networks.
